Skip to contentSkip to conversation

mm/rmap: fix potential batched TLB flush race

PATCH In review Huang Ying 2 participants 2 messages Nov 25, 2021

Subsystems (3)
lore
Share & tools
lore

Huang Ying

author
#0 lore raw

PATCH 1/1 mm/rmap: fix potential batched TLB flush race

In theory, the following race is possible for batched TLB flushing.

CPU0 CPU1
---- ----
shrink_page_list()

                                   unmap
                                     zap_pte_range()
                                       flush_tlb_batched_pending()
                                         flush_tlb_mm()
  try_to_unmap()
    set_tlb_ubc_flush_pending()
      mm->tlb_flush_batched = true
                                         mm->tlb_flush_batched = false

After the TLB is flushed on CPU1 via flush_tlb_mm() and before mm->tlb_flush_batched is set to false, some PTE is unmapped on CPU0 and the TLB flushing is pended. Then the pended TLB flushing will be lost. Although both set_tlb_ubc_flush_pending() and flush_tlb_batched_pending() are called with PTL locked, different PTL instances may be used.

Because the race window is really small, and the lost TLB flushing will cause problem only if a TLB entry is inserted before the unmapping in the race window, the race is only theoretical. But the fix is simple and cheap too.

Syzbot has reported this too as follows,

==================================================================

BUG: KCSAN: data-race in flush_tlb_batched_pending / try_to_unmap_one

write to 0xffff8881072cfbbc of 1 bytes by task 17406 on cpu 1:

 flush_tlb_batched_pending+0x5f/0x80 mm/rmap.c:691
 madvise_free_pte_range+0xee/0x7d0 mm/madvise.c:594
 walk_pmd_range mm/pagewalk.c:128 [inline]
 walk_pud_range mm/pagewalk.c:205 [inline]
 walk_p4d_range mm/pagewalk.c:240 [inline]
 walk_pgd_range mm/pagewalk.c:277 [inline]
 __walk_page_range+0x981/0x1160 mm/pagewalk.c:379
 walk_page_range+0x131/0x300 mm/pagewalk.c:475
 madvise_free_single_vma mm/madvise.c:734 [inline]
 madvise_dontneed_free mm/madvise.c:822 [inline]
 madvise_vma mm/madvise.c:996 [inline]
 do_madvise+0xe4a/0x1140 mm/madvise.c:1202
 __do_sys_madvise mm/madvise.c:1228 [inline]
 __se_sys_madvise mm/madvise.c:1226 [inline]
 __x64_sys_madvise+0x5d/0x70 mm/madvise.c:1226
 do_syscall_x64 arch/x86/entry/common.c:50 [inline]
 do_syscall_64+0x44/0xd0 arch/x86/entry/common.c:80
 entry_SYSCALL_64_after_hwframe+0x44/0xae

write to 0xffff8881072cfbbc of 1 bytes by task 71 on cpu 0:

 set_tlb_ubc_flush_pending mm/rmap.c:636 [inline]
 try_to_unmap_one+0x60e/0x1220 mm/rmap.c:1515
 rmap_walk_anon+0x2fb/0x470 mm/rmap.c:2301
 try_to_unmap+0xec/0x110
 shrink_page_list+0xe91/0x2620 mm/vmscan.c:1719
 shrink_inactive_list+0x3fb/0x730 mm/vmscan.c:2394
 shrink_list mm/vmscan.c:2621 [inline]
 shrink_lruvec+0x3c9/0x710 mm/vmscan.c:2940
 shrink_node_memcgs+0x23e/0x410 mm/vmscan.c:3129
 shrink_node+0x8f6/0x1190 mm/vmscan.c:3252
 kswapd_shrink_node mm/vmscan.c:4022 [inline]
 balance_pgdat+0x702/0xd30 mm/vmscan.c:4213
 kswapd+0x200/0x340 mm/vmscan.c:4473
 kthread+0x2c7/0x2e0 kernel/kthread.c:327
 ret_from_fork+0x1f/0x30

value changed: 0x01 -> 0x00

Reported by Kernel Concurrency Sanitizer on:

CPU: 0 PID: 71 Comm: kswapd0 Not tainted 5.16.0-rc1-syzkaller #0
Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 01/01/2011
==================================================================
Reported-by [email protected]
Cc Nadav Amit <[email protected]>
Cc Mel Gorman <[email protected]>
Cc Andrea Arcangeli <[email protected]>
Cc Andy Lutomirski <[email protected]>
Cc Dave Hansen <[email protected]>
Cc Will Deacon <[email protected]>
Cc Yu Zhao <[email protected]>
Cc Marco Elver <[email protected]>
---
 include/linux/mm_types.h |  2 +-
 mm/rmap.c                | 30 +++++++++++++++++++++++-------
 2 files changed, 24 insertions(+), 8 deletions(-)
include/linux/mm_types.h +1 -1
index c3a6e6209600..789778067db9 100644
@@ -632,7 +632,7 @@ struct mm_struct {
 		atomic_t tlb_flush_pending;
 #ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
 		/* See flush_tlb_batched_pending() */
-		bool tlb_flush_batched;
+		atomic_t tlb_flush_batched;
 #endif
 		struct uprobes_state uprobes_state;
 #ifdef CONFIG_PREEMPT_RT
mm/rmap.c +23 -7
index 163ac4e6bcee..2e6b19be5a18 100644
@@ -621,6 +621,18 @@ void try_to_unmap_flush_dirty(void)
 		try_to_unmap_flush();
 }
 
+/* The upper 15 bit of mm->tlb_flush_batched records pending flushes */
+#define TLB_FLUSH_BATCH_PENDING_SHIFT	16
+#define TLB_FLUSH_BATCH_COUNT_MASK	0x7f
+#define TLB_FLUSH_BATCH_PENDING_ONE	(1 << TLB_FLUSH_BATCH_PENDING_SHIFT)
+
+#define TLB_FLUSH_BATCH_PENDING(cnt)					\
+	(((cnt) >> TLB_FLUSH_BATCH_PENDING_SHIFT) & TLB_FLUSH_BATCH_COUNT_MASK)
+#define TLB_FLUSH_BATCH_FLUSHED(cnt)					\
+	((cnt) & TLB_FLUSH_BATCH_COUNT_MASK)
+#define TLB_FLUSH_BATCH_PACK(pending, flushed)				\
+	(((pending) << TLB_FLUSH_BATCH_PENDING_SHIFT) | (flushed))
+
 static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
 {
 	struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
@@ -633,7 +645,7 @@ static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
 	 * before the PTE is cleared.
 	 */
 	barrier();
-	mm->tlb_flush_batched = true;
+	atomic_add(TLB_FLUSH_BATCH_PENDING_ONE, &mm->tlb_flush_batched);
 
 	/*
 	 * If the PTE was dirty then it's best to assume it's writable. The
@@ -680,15 +692,19 @@ static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
  */
 void flush_tlb_batched_pending(struct mm_struct *mm)
 {
-	if (data_race(mm->tlb_flush_batched)) {
-		flush_tlb_mm(mm);
+	int batched = atomic_read(&mm->tlb_flush_batched);
+	int pending = TLB_FLUSH_BATCH_PENDING(batched);
+	int flushed = TLB_FLUSH_BATCH_FLUSHED(batched);
 
+	if (pending != flushed) {
+		flush_tlb_mm(mm);
 		/*
-		 * Do not allow the compiler to re-order the clearing of
-		 * tlb_flush_batched before the tlb is flushed.
+		 * If the new TLB flushing is pended during flushing,
+		 * leave mm->tlb_flush_batched as is, to avoid to lose
+		 * flushing.
 		 */
-		barrier();
-		mm->tlb_flush_batched = false;
+		atomic_cmpxchg(&mm->tlb_flush_batched, batched,
+			       TLB_FLUSH_BATCH_PACK(pending, pending));
 	}
 }
 #else
Signature · 1 line
2.30.2
#1 lore raw
Quoted 112 lines · Huang Ying

In theory, the following race is possible for batched TLB flushing.

CPU0 CPU1
---- ----
shrink_page_list()

                                  unmap
                                    zap_pte_range()
                                      flush_tlb_batched_pending()
                                        flush_tlb_mm()
 try_to_unmap()
   set_tlb_ubc_flush_pending()
     mm->tlb_flush_batched = true
                                        mm->tlb_flush_batched = false

After the TLB is flushed on CPU1 via flush_tlb_mm() and before mm->tlb_flush_batched is set to false, some PTE is unmapped on CPU0 and the TLB flushing is pended. Then the pended TLB flushing will be lost. Although both set_tlb_ubc_flush_pending() and flush_tlb_batched_pending() are called with PTL locked, different PTL instances may be used.

Because the race window is really small, and the lost TLB flushing will cause problem only if a TLB entry is inserted before the unmapping in the race window, the race is only theoretical. But the fix is simple and cheap too.

Syzbot has reported this too as follows,

==================================================================

BUG: KCSAN: data-race in flush_tlb_batched_pending / try_to_unmap_one

write to 0xffff8881072cfbbc of 1 bytes by task 17406 on cpu 1:

flush_tlb_batched_pending+0x5f/0x80 mm/rmap.c:691
madvise_free_pte_range+0xee/0x7d0 mm/madvise.c:594
walk_pmd_range mm/pagewalk.c:128 [inline]
walk_pud_range mm/pagewalk.c:205 [inline]
walk_p4d_range mm/pagewalk.c:240 [inline]
walk_pgd_range mm/pagewalk.c:277 [inline]
__walk_page_range+0x981/0x1160 mm/pagewalk.c:379
walk_page_range+0x131/0x300 mm/pagewalk.c:475
madvise_free_single_vma mm/madvise.c:734 [inline]
madvise_dontneed_free mm/madvise.c:822 [inline]
madvise_vma mm/madvise.c:996 [inline]
do_madvise+0xe4a/0x1140 mm/madvise.c:1202
__do_sys_madvise mm/madvise.c:1228 [inline]
__se_sys_madvise mm/madvise.c:1226 [inline]
__x64_sys_madvise+0x5d/0x70 mm/madvise.c:1226
do_syscall_x64 arch/x86/entry/common.c:50 [inline]
do_syscall_64+0x44/0xd0 arch/x86/entry/common.c:80
entry_SYSCALL_64_after_hwframe+0x44/0xae

write to 0xffff8881072cfbbc of 1 bytes by task 71 on cpu 0:

set_tlb_ubc_flush_pending mm/rmap.c:636 [inline]
try_to_unmap_one+0x60e/0x1220 mm/rmap.c:1515
rmap_walk_anon+0x2fb/0x470 mm/rmap.c:2301
try_to_unmap+0xec/0x110
shrink_page_list+0xe91/0x2620 mm/vmscan.c:1719
shrink_inactive_list+0x3fb/0x730 mm/vmscan.c:2394
shrink_list mm/vmscan.c:2621 [inline]
shrink_lruvec+0x3c9/0x710 mm/vmscan.c:2940
shrink_node_memcgs+0x23e/0x410 mm/vmscan.c:3129
shrink_node+0x8f6/0x1190 mm/vmscan.c:3252
kswapd_shrink_node mm/vmscan.c:4022 [inline]
balance_pgdat+0x702/0xd30 mm/vmscan.c:4213
kswapd+0x200/0x340 mm/vmscan.c:4473
kthread+0x2c7/0x2e0 kernel/kthread.c:327
ret_from_fork+0x1f/0x30

value changed: 0x01 -> 0x00

Reported by Kernel Concurrency Sanitizer on:

CPU: 0 PID: 71 Comm: kswapd0 Not tainted 5.16.0-rc1-syzkaller #0
Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 01/01/2011
==================================================================
Reported-by [email protected]
Cc Nadav Amit <[email protected]>
Cc Mel Gorman <[email protected]>
Cc Andrea Arcangeli <[email protected]>
Cc Andy Lutomirski <[email protected]>
Cc Dave Hansen <[email protected]>
Cc Will Deacon <[email protected]>
Cc Yu Zhao <[email protected]>
Cc Marco Elver <[email protected]>

---
include/linux/mm_types.h | 2 +-
mm/rmap.c | 30 +++++++++++++++++++++++-------

2 files changed, 24 insertions(+), 8 deletions(-)
include/linux/mm_types.h +1 -1
index c3a6e6209600..789778067db9 100644
@@ -632,7 +632,7 @@ struct mm_struct {
 		atomic_t tlb_flush_pending;
 #ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
 		/* See flush_tlb_batched_pending() */
-		bool tlb_flush_batched;
+		atomic_t tlb_flush_batched;
 #endif
 		struct uprobes_state uprobes_state;
 #ifdef CONFIG_PREEMPT_RT
mm/rmap.c +0 -0
index 163ac4e6bcee..2e6b19be5a18 100644

mm/rmap.c void try_to_unmap_flush_dirty(void) · line 624

Quoted 4 lines
mm/rmap.c +1 -0
@@ -621,6 +621,18 @@ void try_to_unmap_flush_dirty(void)
 		try_to_unmap_flush();
 }
 
+/* The upper 15 bit of mm->tlb_flush_batched records pending flushes */

Why 15? I think it will be easier to swallow if it was 32-bit (which correspond to number of PIDs?)

What would happen on an overflow? If you regarded each pneding/flushed counter as a separate atomic, that would have been easier. But anyhow, a comment is necessary IMHO.

mm/rmap.c void try_to_unmap_flush_dirty(void) · line 626

Quoted 2 lines
mm/rmap.c +2 -0
+#define TLB_FLUSH_BATCH_PENDING_SHIFT	16
+#define TLB_FLUSH_BATCH_COUNT_MASK	0x7f

0x7f is not 15 bits the last time I checked.

mm/rmap.c void try_to_unmap_flush_dirty(void) · line 634

Quoted 8 lines
mm/rmap.c +8 -0
+#define TLB_FLUSH_BATCH_PENDING_ONE	(1 << TLB_FLUSH_BATCH_PENDING_SHIFT)
+
+#define TLB_FLUSH_BATCH_PENDING(cnt)					\
+	(((cnt) >> TLB_FLUSH_BATCH_PENDING_SHIFT) & TLB_FLUSH_BATCH_COUNT_MASK)
+#define TLB_FLUSH_BATCH_FLUSHED(cnt)					\
+	((cnt) & TLB_FLUSH_BATCH_COUNT_MASK)
+#define TLB_FLUSH_BATCH_PACK(pending, flushed)				\
+	(((pending) << TLB_FLUSH_BATCH_PENDING_SHIFT) | (flushed))

I would have preferred, when possible to avoid such macros. It just makes reading the code harder.

mm/rmap.c static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable) · line 648

Quoted 10 lines
mm/rmap.c +2 -1
+
 static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
 {
 	struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
@@ -633,7 +645,7 @@ static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
 	 * before the PTE is cleared.
 	 */
 	barrier();
-	mm->tlb_flush_batched = true;
+	atomic_add(TLB_FLUSH_BATCH_PENDING_ONE, &mm->tlb_flush_batched);

Any reason to put the pending in the top bits instead of the low ones? It can at least simplify the code a bit. As for the barrier, I would change it for smp_mb__before_atomic() or smp_wmb(). You want the PTE changes and the counter updates to be ordered.

mm/rmap.c static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags) · line 684

Quoted 9 lines
mm/rmap.c +0 -2
 	/*
 	 * If the PTE was dirty then it's best to assume it's writable. The
@@ -680,15 +692,19 @@ static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
 */
 void flush_tlb_batched_pending(struct mm_struct *mm)
 {
-	if (data_race(mm->tlb_flush_batched)) {
-		flush_tlb_mm(mm);

The previous smp_mb__before_atomic() or smp_wmb() should be matched with an smp_mb__before_atomic() or smp_rmb() here, I think.

mm/rmap.c static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags) · line 697

Quoted 3 lines
mm/rmap.c +3 -0
+	int batched = atomic_read(&mm->tlb_flush_batched);
+	int pending = TLB_FLUSH_BATCH_PENDING(batched);
+	int flushed = TLB_FLUSH_BATCH_FLUSHED(batched);

I would prefer them being unsigned.

Quoted 5 lines
mm/rmap.c +2 -1
+	if (pending != flushed) {
+		flush_tlb_mm(mm);
 		/*
-		 * Do not allow the compiler to re-order the clearing of

mm/rmap.c static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags) · line 708

Quoted 10 lines
mm/rmap.c +5 -3
-		 * tlb_flush_batched before the tlb is flushed.
+		 * If the new TLB flushing is pended during flushing,
+		 * leave mm->tlb_flush_batched as is, to avoid to lose
+		 * flushing.
 		 */
-		barrier();
-		mm->tlb_flush_batched = false;
+		atomic_cmpxchg(&mm->tlb_flush_batched, batched,
+			       TLB_FLUSH_BATCH_PACK(pending, pending));
 	}

Overall, I am not overly excited about the fact the the mm generation and this batching mechanism remain separated. This makes reasoning about TLB flushes harder and can lead to unnecessary flushes.

I understand the complexity in changing the code to get there, and perhaps your approach is reasonable for now, if you at least manage to deal with overflows.

Thanks for doing all of that, sorry for being negative. I guess I suffer from NIH ("not invented here”).